So why use AI personas as alignment schemes?
First, a mental model: AGENTS md is basically the system prompt for your repo. It’s the first thing a coding agent reads when it walks into your project. It’s your “here’s what matters here,” your entry point, your attempt at orientation before it starts rummaging through the codebase like a raccoon with a flashlight.
Conventional wisdom says: stuff AGENTS md with rules. The more rules, the more control. Add rules when the agent screws up. Add more rules when it screws up again. Eventually you’re running a whole repo with a tiny constitution written by someone who has never seen sunlight.
And that, my children, is where slop comes from.
The reason is simple and mechanical. We now have a pretty clean convergence of research pointing at the same failure mode: high instruction density makes agents worse.
Instruction-following performance degrades as you pile on requirements, and “conflicting / tension-y instructions” are a big part of why it degrades.
Now zoom in from general instruction overload to the exact thing we do in agentic coding: repo context files like AGENTS md.
There’s fresh work showing that “do what the tooling recommends, generate a big context file” often reduces task success rates compared to giving no repo context at all, while increasing inference cost by 20%+ and pushing agents into broader exploration patterns.
Sometimes, respecting the file is worse than ignoring it.
Another recent study finds AGENTS md can reduce runtime and output tokens in PR workflows without changing completion behavior much. So it’s not “AGENTS md bad,” but rather, “instruction mass matters, and unnecessary requirements are a tax.”
Beliefs and boundaries
That’s where personas come in.
Rule stacks are downstream. Beliefs are upstream.
Instead of writing AGENTS md, or any agentic prompting like a courtroom transcript, write it like a clean doorway: jurisdiction, a few beliefs that generalize well, and boundaries that prevent the known spirals.
When you bundle beliefs + boundaries, you get a persona. Structurally, what you’ve built is a lightweight alignment layer that avoids the branch explosion without turning your repo prompt into a tax code.
And that’s my TED talk.
Tip your bartender on the way out.
They’re doing branch pruning in real time too.
Love ya.
Sources:
- “How Many Instructions Can LLMs Follow at Once?” (IFScale), arXiv:2507.11538
- “Boosting Instruction Following at Scale” (conflict scoring / instruction tension), arXiv:2510.14842
- “Evaluating AGENTS.md: Are Repository-Level Context Files Helpful for Coding Agents?”, arXiv:2602.11988
- “On the Impact of AGENTS.md Files on the Efficiency of AI Coding Agents”, arXiv:2601.20404